docs: sync llms.txt with audio input support and agent.embed()

[LeonSGP43]

What

Project Xander (b5e2943) added audio input to agent.run/agent.complete and a public agent.embed() on the Needle 3 retrieval head, and updated doc/apis.md accordingly — but llms.txt still listed the old signatures and omitted agent.embed() entirely, while promising readers that it is enough to write correct Needle code without inventing API.

This PR syncs llms.txt with the shipped code:

• agent.run / agent.complete: document the new text="" default and the audio / audio_format="wav" / sample_rate=0 / channels=1 argument group (Needle 3 models only)
• add the missing agent.embed(text="", audio=None, audio_format="wav", sample_rate=0, channels=1) -> list[float] entry, including the Needle-3-only constraint (raises ValueError on a Needle 2 model)
• add a short audio input note (WAV path or WAV bytes; raw PCM16/float32 buffers pass audio_format plus positive sample_rate/channels), mirroring doc/apis.md

Verification

• Signatures cross-checked against needle/__init__.py on main (complete L215, run L273, embed L254; _prepare_audio L92 accepts wav/pcm16/float32, PCM requires positive rate/channels; generation < 3 raises for audio and embeddings)
• Wording mirrors the doc/apis.md table updated in b5e2943
• Complements docs: sync llms.txt with multi-generation engine changes #107 (multi-generation engine sync): that PR touches the extract line and the CLI section; this PR touches the run/complete lines, adds the embed entry, and appends an audio note after the extraction bullet — no overlapping hunks

Follows the precedent of #99 (llms.txt documentation sync).